/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x8c16-minmax-neon-mlal-padal.c | 183 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 226 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 159 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 202 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 260 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 303 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 190 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 239 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 222 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 271 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 317 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 366 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 247 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 328 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 295 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 376 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 366 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 459 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 302 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 395 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 436 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 517 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 541 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 634 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x8c8-minmax-neon-mull-padal.c | 177 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 220 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 201 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 244 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 210 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 259 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 242 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 291 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 278 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 321 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 337 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 386 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 265 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 346 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 313 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 394 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 322 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 415 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 386 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 479 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 454 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 535 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 561 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 654 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|